Tactile Games

Strategies for Using Proximal Policy Optimization in Mobile Puzzle Games

Automating play-testing in mobile puzzle games requires overcoming significant technical hurdles associated with reinforcement learning, particularly the instability and overfitting inherent in Proximal Policy Optimization (PPO).

• The CSResetMaskV2 framework is the most effective configuration for training PPO agents, achieving the highest completion rates and human-like gameplay patterns in match-three titles like Lily’s Garden.
• Integrating action masks directly into the observation space as a partial forward model significantly improves training speed and reliability compared to traditional hard action masking, which often causes algorithmic collapse.
• To prevent catastrophic learning failures and suboptimal loops, developers should utilize a combination of environment resets, color shuffling, and soft action masking.